A0347
Title: Feature extraction from satellite data for multivariate time-series forecasting of biotoxin contamination in shellfish
Authors: Sergio Tavares - NOVA Laboratory for Computer Science and Informatics - NOVA LINCS / NOVA School of Science and Technology (Portugal) [presenting]
Ludwig Krippahl - NOVA Laboratory for Computer Science and Informatics - NOVA LINCS (Portugal)
Marta Lopes - Department of Mathematics NOVA School of Science and Technology (Portugal)
Abstract: Shellfish production is an important economic activity in Portugal, making shellfish contamination with biotoxins a public health problem and a significant economic risk. Predicting shellfish contamination could improve production management and public health protection. Several years of satellite images obtained from Sentinel-3 mission for marine observation and biotoxin contamination data from shellfish species collected by the Portuguese official control from 16 locations on the western coast of Portugal are used. The goal is to predict when toxin concentration in shellfish will exceed safety limits. The problem is formulated as a time-series forecasting problem, taking as variable past values of contamination and a time series of satellite images for the given locations. Images are available several times a week, while measurements take place once a week. Since images are high-dimensional data, first, a small number of relevant features must be extracted. We do this in an unsupervised manner using autoencoders that are also capable of ignoring non-valid pixels. These frequently occur due to clouds, land or different anomalies. Our results show that including these features improves the prediction of contamination events for 2021 in models trained on data from previous years, showing that with this approach, we can include information from a high-dimension data source like remote sensing without losing the ability of the model to generalize outside the training set.
